The impact of sampling method on serialization effect

[Rfdfz]

Hi, if we use FPS instead of Grid Sample for the point cloud in the data preprocessing stage, what impact will it have on the serialization effect? I followed the annotations below in the code. It works, but I'm not sure if the point cloud is being serialized correctly.

# please add the following augmentation into your pipline:
# dict(type="Copy", keys_dict={"grid_size": 0.01}),
# (adjust `grid_size` to what your want)

[Gofinge]

Good question! In my opinion, it works with FPS, yet tuning a good initial grid size (no need for grid sampling, just copy it into data dict) is still necessary. In that case, multiple points will exist in one single grid, but it won't affect serialization (and SpConv also supports such a pattern, MinkEngine does not support it).

[Rfdfz]

Thanks for your answer!!!